customers
All customer stories
Top Voice AI companies are building with Assembly.
resources
Latest Release
Universal 3.5 Pro
Our new flagship speech-to-text model handles real-world audio the way it actually happens. Available for real-time and pre-recorded audio.
resources
A production-ready framework for evaluating what your agent actually hears. Now independently verified: three benchmarks — Coval, Daily's Pipecat, and Hugging Face — tested Universal-3.5 Pro on their own audio and found leading accuracy at competitive latency.
Coval STT benchmark · Contamination-resistant dataset
Word error rate, lower is better. Universal-3.5 Pro posts the lowest WER of 28 models pooled across Coval's datasets, where the field averages 6.7%. Source: Coval STT benchmark.
Overall WER
3.2%
Lowest WER of 28 models on Coval's contamination-resistant benchmark, pooled across datasets.
Source: Coval STT benchmark
Semantic WER
1.22%
On Daily's Pipecat Pareto frontier at a 282ms median time-to-final.
Source: Daily's Pipecat STT benchmark
P95 latency
354ms
Holds the accuracy/latency frontier even at the worst-case tail, where rivals drop off.
Source: Daily's Pipecat STT benchmark
Context carryover
10.2%
Further WER reduction when you pass the agent's question as context (agent_context).
Source: Universal-3.5 Pro Realtime
Customer story
“We were searching for the best realtime ASR model for our voice agent pipeline in Fireflies. The new Universal-3.5 Pro speech model from Assembly is best so far in terms of accuracy, latency and language switching.”
— Foysal Osmany, Software Engineer at Fireflies
The proof
Every vendor picks the datasets that flatter its model. So the results that matter are the ones we don't run. Over a single week, three independent parties tested Universal-3.5 Pro on their own audio — from clean scripted speech to messy real agent calls — and landed in the same place.
Coval
3.2%
Rebuilt in July 2026 on a fresh 3,500-sample dataset across six real-world conditions — clipping, far-field mics, phone-codec reverb, accents, and noise. Universal-3.5 Pro posted the lowest overall word error rate of any model pooled across datasets, and topped the field on clipping and far-field.
Daily's Pipecat
1.22% @ 282ms
1,000 real agent conversations scored on semantic WER against time-to-final. Universal-3.5 Pro Realtime sits on the Pareto frontier and holds it at the 354ms P95 tail, where Deepgram falls off. The only way to beat it on accuracy is to accept a slower model.
Hugging Face
Near top
The community-run leaderboard for clean, scripted speech — the opposite end from chaotic agent audio. Universal-3.5 Pro placed near the top there too. Leading on both ends — pristine and messy — is the combination that is hard to fake.
The framework
Generic WER tells you whether the system got words right. These tell you whether your voice agent will actually complete tasks — and whether it stays fast when it does. A serious evaluation measures all of them, not a single headline number.
The metric that predicts production success: can the model transcribe emails, phone numbers, and proper nouns in messy real audio? Measured as word/entity error rate on real agent conversations.
Universal-3.5 Pro RealtimeDoes the model use what it just asked? Pass the question with agent_context and a mumbled reply resolves to "user@assemblyai.com," not "user at assembly a i dot com." No other streaming STT offers this.
Universal-3.5 Pro RealtimeMeasure time-to-final at the median and the P95 tail — the occasional slow turn is what breaks conversational flow. Universal-3.5 Pro posts 282ms median / 354ms P95 and exposes three modes: min_latency, balanced, max_accuracy.
Universal-3.5 Pro RealtimeBad turn detection means the agent cuts you off or sits in dead air. Speech-aware end-of-turn detection reads tonality, pacing, and rhythm — not just silence — and lands around 300ms.
Voice Agent APISession resumption
What happens when the connection drops? In production, a dropped connection shouldn't mean a dropped call. AssemblyAI's Voice Agent API reconnects within 30 seconds with full context preserved.
In production telephony, connections drop — mobile networks switch towers, Wi-Fi flickers, load balancers cycle. A fragile session restarts the whole interaction.
The Voice Agent API holds the session open for 30 seconds. Reconnect inside that window and the conversation resumes exactly where it left off.
System prompt, conversation history, and tool state carry across the reconnect — the caller never notices, and the agent never loses the thread.
The paths
AssemblyAI offers two paths to building voice agents, both built on the same Universal-3.5 Pro Realtime foundation — so transcription quality and accuracy are consistent regardless of which you choose.
Recommended
Evaluate the full pipeline end-to-end
$4.50 /hr
STT + LLM + TTS included · free tier available
Bring your own stack
Evaluate transcription accuracy and latency in isolation
$0.45 /hr
Transcription only · unlimited concurrent streams
The checklist
Running a POC that actually predicts production performance requires testing the right things. Most teams over-test on easy cases and under-test on the edge cases — and the latency tail — that determine real-world success.
POC scorecard
Provider evaluation
Test with your actual audio, not clean samples. Real recordings — background noise, accents, connection quality. Enable voice_focus for noisy or far-field audio.
Test entity-heavy utterances specifically. Emails, phone numbers, alphanumeric codes, proper nouns. This is where agents lose users.
Test whether agent_context helps. Pass the agent's question in the audio and measure the short-utterance accuracy gain.
Measure P95 latency under load, not just the median. The occasional slow turn is what users feel — reproduce the frontier on your own audio.
Test interruptions and barge-in. Start talking while the agent responds. Does it stop and listen? Does it lose context?
Test session drops and reconnection. Kill the WebSocket. Can you reconnect and pick up where you left off within 30 seconds?
Benchmarks are a starting point, not the finish line. Run your own conversations through Universal-3.5 Pro Realtime.
